Research On WLAN Indoor Positioning Algorithm Based On Access Point Selection And Fingerprint Expansion

In the era of mobile Internet,many commercial applications are inseparable from the accurate acquisition of people's location information.In the outdoor wide-area space,the Global Navigation Satellite System(GNSS),represented by GPS,is the most effective means of positioning.Since satellite signals are blocked by buildings,GNSS does not work well indoors.A variety of indoor positioning technologies has emerged,but there is no one recognized standard means.With the popularity of WLAN and smart phones,WLAN fingerprint positioning has become a widely used indoor positioning means.In order to reduce the system computational consumpution and labor cost as well as improve the positioning accuracy and operational efficiency,this paper focuses on the improvement of the positioning model in WLAN fingerprint positioning and the related strategies for reducing the fingerprint collection workload.The research contents include: Access Points selection algorithm of LocalReliefF-C,reference locations clustering algorithm of CSBA,location estimation algorithm of Hidden Na?ve Bayes,the automatic expansion framework of fingerprint database based on GPR and CGAN,the positioning algorithm based on RSS plane search and superposition,the implicit crowdsourcing fingerprint acquisition and positioning system framework,the fingerprint preprocessing method of min-max and the semi-supervised positioning algorithm of Co-RandomForest.(1)The number of WLAN APs in the indoor environment is increasing,which increases the computational complexity of the fingerprint positioning system.Besides,there are some redundant APs and noisy APs that can be removed.In order to select the best location-discriminant APs set,an AP selection algorithm LocalReliefF-C based on location discriminant capability accessment and redundant feature detection is propsosed.Firstly,the algorithm uses an improved feature selection method ReliefF to calculate the classification weight of each AP in the fingerprint sample set that measures the location discriminant capability of the AP and selects several APs with the largest weight values.Subsequently,the algorithm calculates the maximum information coefficient(MIC)between each pair of selected APs to pick out the the redundant APs and removes them.The experimental results show that the AP selection algorithm significantly reduces the computational cost of the positioning algorithm,and the selected set of best discriminant APs ensures good positioning accuracy.(2)In order to search for the target location,the traditional positioning algorithm traverses the fingerprint samples of each reference location,and the search efficiency decreases drastically as the number of reference locations increases.In order to improve the searching efficiency,a new reference location-clustering algorithm CSBA is proposed based on the previous AP selection results.On the offline stage,all the reference locations are divided into several clusters to realize the localization of the search space.On the online stage,the target cluster is determined first and then the target location within the cluster is determined.The criterion to divide multiple reference locations into one cluster is that their best location-discriminant APs set has a certain number of common elements.The experimental results show that the CSBA algorithm effectively reduces the matching times during the positioning process and greatly improves the efficiency of location searching.(3)During the process of estimating the location within the cluster,an improved positioning model,Hidden Na?ve Bayes(HNB),is introduced,which breaks the idealized assumption of the classic Naive Bayes(NB)algorithm on the conditional independence between Access Points.By defining a hidden parent node for each AP,HNB incorporates the interaction between Access Points into the location estimation process.Experiment results show that the HNB algorithm achieves better positioning accuracy than the classic NB algorithm.(4)Under the condition that the number of sampled reference locations is limited and the number of fingerprint samples is insufficient,the positioning performance is difficult to guarantee.Under the premise of not increasing the workload of sample collection,an automatic generation and expansion strategy of fingerprint database is proposed.Firstly,the Gaussian Process Regression(GPR)is used to model the relationship between sample RSS values and location coordinates,which helps to generate fingerprint samples on unsampled reference locations.Then,the Conditional Generative Adversarial Networks(CGAN)can generate a large number of fingerprint samples that are very similar to the existing samples,which further enriches the fingerprint database.(5)Based on the above work,a positioning algorithm is proposed in the location estimation stage based on AP's RSS planes search and superposition.Firstly,the matrix plane of RSS values for each AP on the positioning area is constructed.For a new RSS observation vector,the target location area for each AP are respectively found by searching in its corresponding RSS plane,and finally the result location is determined by superimposing multiple target areas.(6)The traditional fingerprint collection process requires expensive time and labor costs,which becomes a bottleneck for the popularization of fingerprint positioning systems.A system framework of implicit crowdsourcing fingerprint acquisition and location estimation is proposed,which includes several core components such as the sample collection client,the preprocessing module of fingerprint samples and the location estimation module of semi-supervised learning.The crowdsourcing model distributes heavy sample collection tasks to a large number of volunteers,and the fingerprint server implements receiving and integration of uploaded samples.Requiring no user's intervention,the collection program runs silently as a background process of the device,which further improves the practicability of the system.The min-max normalization method implements the preprocessing of collected fingerprints,maps all the RSS values to the same value range,and discretizes them,which effectively solves the problem of observation device heterogeneity during the crowdsourcing fingerprints collection and prepares data for the subsequent location estimation process.(7)A large number of unlabeled samples are collected by implicit crowdsourcing volunteers.In order to use them for positioning,a semi-supervised learning positioning algorithm,Co-RandomForest,which fuses collaborative training and ensemble learning ideas,is introduced.The algorithm optimizes the random forest classifier using the unlabeled samples set in a collaborative training manner,and takes advantage of the ensemble learning to simplify the process of determining the most confident samples.The experimental results show that Co-RandomForest can obtain the comparable positioning performance as the traditional supervised learning algorithm by leveraging a large number of easily obtained unlabeled samples as long as a small number of labeled samples,which thus eliminates the dependence on collecting a large number of labeled samples.